Jul 2026 · The Watch
Open source isn’t automatically cheaper.
“Open-weights means cheap and efficient” is one of the most repeated lines in AI — and the newest heavy open model quietly breaks it on threeaxes at once. If you’re choosing a model to save money, this is the honest math nobody runs for you.
The honest counterweight to our own DeepSeek piece
We cheered DeepSeekfor making serving ~7× cheaper. A careless reading of that is “open source is cheap.” It isn’t, as a rule — DeepSeek pulled off a specificserving win; the next heavy open model is the counterexample. Cost varies enormously. Don’t generalize one win into a law.
The real cost of an open model has three multipliers.
Take Moonshot’s new open-weight coding model (weights dropping late July). It’s genuinely strong — near-frontier at coding. It is also the opposite of the “cheap and efficient” stereotype, because total cost is a product, not a price tag:
- Serving footprint. Top performance reportedly needs ~64 accelerators — a corporate installation, not your laptop and not one cheap GPU.
- Per-token price. If you rent it from the cloud instead, it’s priced in frontier territory (~$15 per million output tokens) — not the bargain “open” implies.
- Token efficiency. It burns more tokens to reach the same answer than the terse closed frontier models — so even at a given price, the meter runs longer.
Multiply those together and a “cheap” open model can cost morethan a “pricey” closed one.
The axis nobody quotes you: tokens-to-answer.
This is the hidden one. Everybody compares price per token. Almost nobody compares how many tokens a model spends to finish the job. A terse frontier model that costs 5× per token but uses a third of the tokens can be cheaper overallthan a “budget” model that rambles to the same answer.
Scruuge’s TLDR
Total cost = price per token × tokens-to-answer × serving cost. Sticker price is one of three terms. Judge a model by the finished bill, not the per-token rate.
“Free weights” are not free to serve.
Open weights mean you canrun it yourself. They don’t mean it’s free — you pay either in compute (the GPUs to host a heavy model, plus the engineers to keep them fed) or in a provider’s markup if you rent. And there’s no free lunch on quality either: to chase the frontier, open models have to scale up, which makes them harder and pricierto serve, not easier. “Small, cheap, and frontier-level” is not a combination physics is currently offering.
So when does open source win?
Plenty of times — just rarely for the reason people say. Reach for open weights when you need:
- Control & fine-tuning. Closed models guardrail against fine-tuning and cloning; open weights let you adapt the model to your data and use cases.
- Resilience. A model you hold can’t be deprecated, rate-limited, or blacked out from under you. In a world of surprise outages and possible distribution restrictions, having at least one model plus a backup is a real risk hedge.
- Specific unlocked tasks. Work a closed vendor politely declines (e.g. wholesale cloning a piece of software) an open model will just do.
Those are strategic reasons — control, portability, capability. “It’s cheaper” is often noton the list, and pretending it is leads to a nasty invoice.
Pick on total cost — we’ll do the math.
Don’t choose a model on its sticker price. The calculator scores what a switch actually costs you, and Know Your Models breaks down every axis that moves the bill — including token efficiency. Open or closed, the honest answer is the finished number.
Prompted by the release of Moonshot’s Kimi K3 open-weight model and a widely-shared analysis of it. The serving-footprint, pricing, and token-efficiency observations are the cost-relevant takeaways; the analyst’s broader claims about which labs lead are one view, not settled fact, and we’ve left them there. Scruuge’s contribution is the total-cost frame and the “when open still wins” list.